4.1. Traffic Crash Characteristics in Shenzhen, China
Among all the crash types, ‘collision with motor vehicles in transport’ is the dominant one, accounting for 66.5% of all the crashes, 51.0% of the deaths, 66.2% of the injuries. NHTSA (2018) also reported that this crash type was the most common first harmful event in fatal crashes (39.2% of all fatal crashes) [
55]. These numbers show that ‘collision with motor vehicles in transport’ challenges traffic safety the most and is urgently needed to be solved in the vehicle and transportation safety technology communities. The statistics in the U.S. show that collisions with fixed objects and non-collisions together accounted for 39.0% of all fatal crashes in 2017, and the number for rollover crashes is 17.1%. These numbers in the U.S. are much higher than the numbers in Shenzhen, China, indicating that traffic management countermeasures and intelligent systems in vehicles should be designed differently across countries.
According to the hourly records of weather in Shenzhen during 2014–2016 [
56], the weather was rainy for 26.4% of the time which is greater than the percentage of crashes in rainy weather (i.e., 11.4%). However, it has been frequently reported in previous studies that drivers are more likely to be involved in crashes on rainy days mainly because of the degraded vision field [
27]. This is not in contradiction with the results presented in
Table 5 because people would avoid traveling on rainy days (especially in heavy rain) for safety if applicable [
57]. Therefore, the reduced traveling frequency on rainy days would probably lead to a lower percentage of crashes on rainy days (11.4%) than the percentage of rainy days in the three years (26.4%).
With respect to crash causation, ‘other unsafe driver behavior while driving’ (CC9) ranks the highest, accounting for 53.2% of all the crashes and 58.5% of all the deaths in the three-year dataset of this study. This causation covers driver distraction, drowsy driving, drunk driving, driving on call, pedestrian or cyclist not following traffic rules, etc. However, the exact detailed causations were not recorded by the polices. To further improve the quality of police-reported crash records for driving safety enhancement in Shenzhen, traffic policies should clearly specify the detailed causations in the crash records in the future. Rear-end crashes have been well-reported as one of the most typical crash types, which is complied with the results in this study that ‘not following with a safe distance’ (CC8) ranks the second and accounts for 15.1% of all the crashes. Besides, frequent lane change significantly challenges driving safety. The results presented in
Table 6 show ‘unsafe lane change’ (CC2) leads to 9.8% of all the crashes in the examined crash records. In general, the top three causations (CC9, CC8, and CC2) account for 78.1% of all the crashes. The statistics shown in
Table 6 could provide an overview of the crash causations in Shenzhen and guide the traffic-safety-related studies, policies, and practical countermeasures to address the safety issues.
Usually, people drive to work around 8 a.m. and go back home around 4–7 p.m., resulting in congestions during morning and evening peak hours [
58,
59]. However, the number of crashes that occurred from 7 a.m. to 9 a.m. is far fewer than the numbers during the evening peak hours in this study. This may probably because drivers do not have many negative emotions in the morning, while their emotions would be affected by their experience during the day, resulting in negative (typically depressed and irritable) emotion states while driving in the evening peak hours [
60]. Previous studies have reported that negative emotion states are closely related to aggressive driving which significantly challenges driving safety [
4,
61]. This explains why there are more crashes during the evening peak hours than during the morning peak hours. Differently, the number of crashes in the evening peak hours on Friday was more than the other workdays probably because of the more activities on Friday night [
62]. NHTSA (2018) also reported that Friday night is the deadliest periods throughout 2017 [
55]. Fewer crashes occurred on Saturday and Sunday morning, and the number of crashes on Sunday is the lowest in a week probably because people tend to enjoy their stay at home on Sunday [
62]. The found crash characteristics with respect to the day of the week are consistent with the reported trends in previous studies [
58,
63].
Considering the percentage of responsibility-prone drivers in each age group of the crash-involved drivers, our results show that the number is the lowest for drivers younger than 18 probably because minors would only be allowed to drive under the supervision of their parents and the presence of parents degrades their aggressive driving because of the monitoring effect [
64], while the number reaches the highest for age group 19∼25 mainly because of the no monitoring effect from their parents and drivers’ higher violation rates, underestimation of various violation risks, lower level of motivation to follow traffic rules, and overly involved in running red lights than older mature drivers [
65]. Similarly, young male drivers involved in fatal collisions were twice as likely to be speeding as male driver from the ages of 35 to 44 in the U.S., in 2013 [
66].
Compared with the drivers of passenger cars, professional drivers such as truck and bus drivers spend more time behind the wheel dealing with complicated driving tasks, and they are regulated by higher requirements on transport efficiency and fuel consumption. The higher requirements and long-time driving would lead the drivers to be fatigue and/or distracted, which would increase the probability to be involved in full responsibility crashes. Another reason leading to the higher involvement of truck and bus drivers in full responsibility crashes is that the blind zones of buses and trucks are larger than cars. Because objects (e.g., pedestrians, fixed objects on-road) in the blind zones of a vehicle are difficult to be noticed by the driver [
67], the larger blind zones of buses and trucks would make the blind zone related crash risk higher than cars. This characteristic of bus/truck driving makes it attention-demanding to take care of surrounding road users and keep a safe distance from them. Therefore, ADAS systems for buses and trucks should properly address the safety challenges in the blind zones to improve driving safety, such as the blind zone warning systems.
In exploring the land-use patterns of the occurred crashes, we found that rural and central urban areas are featured with more frequent crashes than the areas with other land-use patterns. A study found that densely populated areas for public services may increase the traffic risks [
68]. This is partly consistent with our finding in this study that the areas of high land-use intensity (LUC5) have the second-highest number of crashes. Interestingly, the highest number of crashes occurred in the rural areas (LUC2) with low land-use intensity in Shenzhen, different from the previous knowledge that rural areas usually have fewer or, at most, a similar number of crashes [
55]. The illustrated results in
Figure 6a show that the LUC2 mainly distributes along the coastline where people in Shenzhen intensively go for walking every day. Typical characteristics of the areas near the coastline in Shenzhen include that there is almost no commercial shops or stores and the around city expressways and main roads are with higher speed limits (usually 60∼80 km/h), resulting in the crash locations being clustered as rural areas. The high density of people activities and the high driving speed in LUC2 may be the leading causes of the high numbers of crashes and deaths. Our statistical analysis results also show that there is no statistical significance between LUC2 and LUC5 in Shenzhen regarding the number of fatalities, but the number of injuries in LUC5 is significantly greater than the number in LUC2. Differently, a meta-analysis on the relationship between speed and road safety [
69] show that lower speed has a more positive effect on reducing fatalities in rural areas than in urban areas, but the lower speed has a more positive effect on reducing injuries in urban areas than in rural areas. According to the findings in [
69], the LUC5 with lower driving speed should have fewer injuries than LUC2, which is different from our findings. This may probably because of the differences between the real rural areas in [
69] and the coastline areas with rural characteristics in Shenzhen, but this needs further and deeper investigations.
In our Bayesian network analysis, we summarize the relationship between the crash attributes where the crash causation is dependent on the crash type and driver responsibility is highlighted. The results suggest that the crashes between motor vehicles and pedestrians are not always due to the misbehavior of the driver side. Even for those crashes where drivers take part of the responsibility, there are also some unsafe behaviors from the pedestrian side. Although the detailed unsafe behaviors of drivers and/or pedestrians were not recorded, one of the major unsafe driver behaviors from the existing literature is drunk driving [
55]. Of the persons who were killed in crashes in 2017 in the U.S., 29% died in alcohol-impaired driving crashes [
55]. This again highlights the importance of specifying and regulating the taxonomy of crash causations in police-reported crash records in Shenzhen, China. Besides, our Bayesian network analysis results show that driver age is found to be associated with vehicle type (driver age → vehicle type). This is intuitive since most bus and truck drivers in China are young or middle-aged drivers. Different vehicle types result in different patterns of crash type and driver responsibilities (vehicle type → crash type, driver responsibility), confirming the results presented in
Section 3.1.8 where we found that a bus or truck driver is more likely to take full responsibility of the crash occurrence than car drivers and motorcycle drivers in bus or truck-involved crashes. Besides, road type and crash causation are also found to be associated with land-use pattern of the occurred crashes. This has been described and discussed in detail in
Section 3.2 For example, LUC5 indicates the high land-use intensity areas where urban roads with low-speed-limit and crash with low-speed vehicles and pedestrians are the typical characteristics.
4.2. Implications of the Findings in This Study
China has an emerging driver population and cultural values that result in aberrant driving behaviors and scrambling to gain the right of way, producing a high number of crashes [
5]. Although it has been frequently reported in previous studies that attributes including driver gender, age, day of the week, time of day, weather, etc. affected drivers’ involvement in crashes [
52,
55,
70,
71], few studies in the literature have investigated the crash characteristics in China from multi-aspects based on policed-reported crash records. Finding out these characteristics would help understand the behavior of Chinese road users and guide the research/application focuses to further improve road safety from the following aspects in general:
(1) Many missing values (the unknown values in the tables) or attributes (e.g., driving exposure information, property loss) and un-clarified details (e.g., the exact causation in the ‘other unsafe driver behavior’ category) can be observed in our presented results. To avoid these problems in future crash records, portable electric devices can be developed and deployed for each traffic police to record crash details with a list of required inputs including the automatically recognized GPS information, the specified unsafe driver behavior (e.g., drunk driving), etc. This would add value to the collected data for traffic characteristics analysis.
(2) Current ADAS systems have been extensively focused on the prevention of collisions with a motor vehicle in transport (CT3). However, sideswipe crashes with pedestrians (CT6) and crushing pedestrian crashes (CT11) have not been well addressed in the current literature. Our crash type results show that CT6 and CT11 crashes accounted for 6.8% and 0.2% of all the crashes but caused 19.4% and 11.1% of all the deaths, respectively. Therefore, pedestrian-related ADAS systems should be well developed to address the high death numbers in pedestrian-related crashes. Besides, the selection of typical scenes is critical for the development of ADAS systems [
72,
73]. The frequently observed crash causations (e.g., not following with a safe distance, unsafe lane change), crash types (e.g., CT3, CT6, CT11), occurrence locations (e.g., normal urban road/street) in this study could be selected as the typical scenes in priority for the development of solutions to improve driving safety in Shenzhen.
(3) Land-use pattern affects crash occurrence. Our results show that most of the crashes happened in rural areas (LUC2) dominated by highway collisions in Shenzhen, different from the second-ranking land-use cluster LUC5 (the areas of high land-use intensity). LUC2 is with low land-use intensity and it mainly distributes along the coastline with speed limits from 60∼80 km/h, while the LUC5 mainly distributes in the central urban areas with extensive human activities. These results indicate that the traffic management authorities in Shenzhen should design different strategies to prevent crashes in the areas with different land-use patterns. For example, more deceleration zones and alerting signals (e.g., flashing lights) can be considered in the areas where crashes frequently occur in LUC2, while more guardrails can be used to regulate pedestrian behavior and to separate motor-vehicles from pedestrians in LUC5.
(4) The Bayesian network analysis reveals the complex interplay between the examined attributes of motor-vehicle crashes, complementing the single-attribute analysis. For example, the interplay results show that vehicle type affects drivers’ responsibility in crashes (vehicle type → driver responsibility). This consolidates the finding in the single-attribute analysis that a bus or truck driver is more likely to take full responsibility for the crash occurrence than the car or motorcycle drivers if involved in crashes. This consolidated finding by the Bayesian network analysis indicates that more ADAS functions should be designed and developed to help bus and truck drivers. For instance, driver fatigue detection and warning functions for truck drivers, blind-zone monitoring and warning functions for both truck and bus drivers (as presented in detail in
Section 4.1), etc.